Intelligent scheduling management method and system for beaming workshop

Abstract

The application belongs to the field of production control, and specifically discloses an intelligent scheduling management method and system for a beaming workshop, which comprises the following steps: firstly, obtaining beaming workshop data such as bobbin task data and beaming machine state data; secondly, establishing a bipartite graph with bobbin task nodes and beaming machine nodes as top points, building edges based on constraints such as color number matching, and generating a feature vector; thirdly, inputting the bipartite graph into a pre-trained graph neural network model, and outputting the matching score of the edges through coding aggregation; fourthly, filtering the edges according to hard constraints, and solving the global matching problem of the weighted bipartite graph to obtain an optimal distribution scheme; and finally, converting the scheme into a device control instruction and executing it. The application realizes global optimal scheduling by integrating multi-dimensional constraints, improves production efficiency and product quality, adapts to dynamic changes in the workshop, and improves the intelligent management of the beaming workshop.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of production control, and particularly relates to an intelligent scheduling management method and system for a beaming workshop. BACKGROUND

[0002] In the production process of a beaming workshop in the textile industry, beaming scheduling is a key link affecting production efficiency and product quality. Traditional beaming scheduling adopts linear matching logic and only allocates cones according to the idle state of beaming machines and the matching degree of speed, without considering the multi-dimensional complexity of beaming tasks, which leads to difficulty in guaranteeing quality consistency and limitation of production efficiency.

[0003] Specifically, a beaming task is affected by multiple factors: cone batches need to be unified to avoid dyeing differences, color numbers need to be consistent to prevent product color differences, tension needs to be within the range allowed by beaming machines to prevent yarn breakage or uneven winding, and at the same time, the path load of the conveying line (to avoid congestion and overtime) and the task priority (to meet the rush demand) need to be considered. The traditional scheduling ignores the above factors, which easily leads to problems such as re-allocation, equipment downtime, color difference exceeding the standard, and increases the cost of rework and production delay.

[0004] In addition, the state of the workshop is dynamic, and the position of the cone, the running state of the beaming machine, and the path load change over time. The traditional static scheduling cannot be adapted in real time, further reducing the rationality of scheduling. SUMMARY

[0005] To solve the above technical problems, the application provides an intelligent scheduling management method and system for a beaming workshop.

[0006] The application provides an intelligent scheduling management method for a beaming workshop, comprising the following method steps: S1, obtaining beaming workshop data, wherein the beaming workshop data comprises cone task data, beaming machine state data, workshop topology and path information, and historical scheduling records.

[0007] S2, establishing a bipartite graph with cone task nodes and beaming machine nodes as vertices, establishing directed edges based on color number matching, path reachability and tension constraints, and generating feature vectors for nodes and edges.

[0008] S3, inputting the bipartite graph into a pre-trained graph neural network model, wherein the graph neural network model encodes and aggregates node and edge information through a node embedding layer and multiple message passing layers, and finally outputs a matching score representing the comprehensive matching advantages and disadvantages of each directed edge through an edge scoring head.

[0009] S4, filtering all the directed edges based on preset hard constraints, and removing edges that do not meet color compatibility, tension range or path reachability.

[0010] S5, taking the filtered directed edges and their matching scores as input, construct and solve a weighted bipartite graph global matching problem aiming to maximize the total matching score, thus obtaining the optimal assignment scheme of the cone tasks to the warping machines.

[0011] S6, convert the assignment scheme into specific device control instructions and issue them to the plant execution system for execution.

[0012] Preferably, the warping plant data is pre-processed, including setting a sliding time window to collect time-series data, normalizing continuous numerical features, numerically encoding or embedding discrete categorical features, and imputing missing data.

[0013] Preferably, the cone task nodes contain normalized speed sequence abstract, remaining cone length, color number embedding vector, tension level, priority and current path load features; the warping machine nodes contain current machine speed, target color number embedding vector, allowed tension range, state code, priority and expected idle time features; the directed edges contain speed matching degree, path transmission delay estimation, historical pairing success rate, color number compatibility score and tension penalty value features.

[0014] Preferably, the bipartite graph is a time-series dynamic graph, which integrates the time-series variation features of the cone tasks and the warping machines into the node and edge feature representations by introducing time encoding or using a sliding time window, to capture the dynamic evolution of the plant state.

[0015] Preferably, the graph neural network model adopts a bidirectional message passing mechanism, including forward message passing from the cone task nodes to the warping machine nodes, and backward message passing from the warping machine nodes to the cone task nodes; the message passing fuses the hidden state features of the connected nodes and the edge features, and uses an attention mechanism to weight and aggregate the messages from different neighbor nodes.

[0016] Preferably, the training of the graph neural network model adopts a multi-objective joint loss function for supervised training; the joint loss function at least includes: a supervised learning loss based on historical successful pairing data, an ordering contrast loss for ensuring that high-quality pairing scores are higher than low-quality pairing, and a constraint-aware loss for imposing a penalty on pairing that violates tension or path capacity constraints.

[0017] Preferably, to solve the weighted bipartite graph global matching problem, the Hungarian algorithm or the minimum cost flow algorithm based on maximum flow is used to obtain the globally optimal matching under the condition that each cone task is assigned at most once and the number of tasks received by each warping machine does not exceed its task slot capacity constraint.

[0018] Preferably, during execution, the system state is monitored in real time and compared with the prediction of the graph neural network, when the deviation exceeds a threshold, a local or global re-planning process is triggered, steps S2 to S5 are re-executed to generate a new scheduling scheme.

[0019] Preferably, the re-planning process is a local re-planning, when an anomaly is detected, only for the subgraph consisting of the affected cone task nodes and the beaming machine nodes, the edge scoring and matching decision are re-performed, while the scheduling scheme of the unaffected part of the system remains unchanged.

[0020] In another aspect, the application also provides an intelligent scheduling management system for a beaming workshop, comprising: a beaming workshop data acquisition module, configured to acquire beaming workshop data, the beaming workshop data comprising cone task data, beaming machine state data, workshop topology and path information, and historical scheduling records.

[0021] A bipartite graph construction module is configured to establish a bipartite graph with cone task nodes and beaming machine nodes as vertices, and to establish directed edges based on color matching, path accessibility and tension constraints, and to generate feature vectors for nodes and edges.

[0022] A matching score module is configured to input the bipartite graph into a pre-trained graph neural network model, wherein the graph neural network model encodes and aggregates node and edge information through a node embedding layer and multiple message passing layers, and finally outputs a matching score representing the overall matching pros and cons for each directed edge through an edge scoring head.

[0023] A directed edge filtering module is configured to filter all the directed edges based on preset hard constraints, and to remove edges that do not meet color compatibility, tension range or path accessibility.

[0024] An allocation scheme acquisition module is configured to input the filtered directed edges and their matching scores as input, to construct and solve a weighted bipartite graph global matching problem with the goal of maximizing the total matching score, and to obtain the optimal cone task to beaming machine allocation scheme.

[0025] An execution module is configured to convert the allocation scheme into specific device control instructions and issue them to the workshop execution system for execution.

[0026] The application realizes global optimal matching of the bobbin-winding machine by modeling the time sequence dynamic bipartite graph with the graph neural network (GNN) as the core, integrating multi-dimensional constraints such as bobbin batch, color number, tension, etc., effectively solving the problem of ignoring quality consistency and multi-factor coupling conflict in traditional scheduling, reducing abnormal risks such as color difference and broken yarn, and reducing rework and downtime costs; the global matching is solved by using the Hungarian algorithm or the minimum cost flow algorithm, taking into account the optimality of small-scale scenes and the real-time performance of large-scale scenes; combined with real-time monitoring and local re-planning, the application dynamically adapts to changes in the workshop state, avoids scheduling lag, and improves the production efficiency and intelligent scheduling level of the warping workshop while ensuring product quality. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor.

[0028] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments with reference to the following drawings.

[0029] Figure 1 A warping workshop intelligent scheduling management method flowchart is provided for the embodiments of the present application.

[0030] Figure 2 A GNN model architecture diagram is provided for the embodiments of the present application.

[0031] Figure 3 A bidirectional message passing flowchart is provided for the embodiments of the present application.

[0032] Figure 4 A local re-planning flowchart is provided for the embodiments of the present application.

[0033] Figure 5 A warping workshop intelligent scheduling management system diagram is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0035] As Figure 1 shown, the embodiment of the application discloses a kind of intelligent scheduling management method 100 of warping workshop, comprising following method steps: S1, obtain warping workshop data, the warping workshop data includes cone yarn task data, warping machine state data, workshop topology and path information and historical scheduling record.

[0036] S2, establish the two-part graph with cone yarn task node and warping machine node as vertex, establish directed edge based on color number matching, path accessibility and tension constraint, and generate feature vector for node and edge.

[0037] S3, the two-part graph is input to a pre-trained graph neural network model, wherein the graph neural network model encodes and aggregates node and edge information through node embedding layer and multiple message passing layers, and finally outputs a matching score representing the comprehensive matching advantages and disadvantages for each directed edge through an edge score head.

[0038] S4, based on the preset hard constraint, filter all the directed edges, and remove the edges that do not satisfy color number compatibility, tension range or path accessibility.

[0039] S5, the filtered directed edge and its matching score are input to construct and solve a weighted two-part graph global matching problem with the goal of maximizing the total matching score, so as to obtain the optimal cone yarn task to warping machine allocation scheme.

[0040] S6, the allocation scheme is converted into specific device control instructions and issued to the workshop execution system for execution.

[0041] In some embodiments, for S1, for cone yarn task data, exemplarily, it is extracted from workshop MES and conveyor PLC real-time database, for example, covering all cone yarn tasks currently in the conveyor line to be allocated, i.e. about to be hung.

[0042] Specifically, it can include basic identification information, for example, batch number (such as A01, A type dyeing batch), color number (such as #FF0033), cone yarn number (unique code, such as TS-001), current conveyor ID (such as Conveyor-03, identifying the current conveyor line where the cone yarn is located).

[0043] Timing and physical characteristics, for example, arrival timestamp (system record time of cone yarn entering the current conveyor line, accurate to seconds), current sampling speed history sequence (within 180-300 seconds, the cone yarn conveying speed data collected by conveyor PLC every 5 seconds forms a timing sequence), remaining cone length (unit: meter, synchronized to PLC from laser length measuring instrument data before cone yarn hanging, updated in real time).

[0044] Process and priority features, such as tension level (e.g. 1-5 levels, 1 level lowest, 5 level highest, tension detection data from previous winding process machine feedback, synchronized to MES), priority label (e.g. “urgent”, “regular”, “low priority”, set in ERP according to order delivery period, synchronized to MES cone yarn task table).

[0045] For the beaming machine state data, exemplarily, collected from the beaming machine local PLC control system, MES and machine HMI (human-machine interface) background database, reflecting the current running state and process requirements of each beaming machine in real time.

[0046] Specifically, it can include basic and state information, such as beaming machine ID (unique code, such as “WJ-08”), current task state (enumeration value: idle / start / run / end yarn / maintenance, determined by beaming machine PLC according to device sensor signals, such as “run” state corresponding to machine spindle speed > 0 and no fault alarm).

[0047] Process parameter information, such as target color number / formula (cone yarn color number and process formula number required by the current order to be processed, such as #FF0033-Recipe05), current machine speed (real-time speed of beaming machine spindle, unit: m / min, collected by PLC speed sensor), allowable tension range (e.g. “20-25N”, pre-set in MES according to process formula, synchronized to beaming machine PLC).

[0048] Scheduling related information, such as priority (corresponding to cone yarn task priority, such as the priority of beaming machine associated with urgent order is set to the highest), estimated remaining processing time (unit: min, calculated by PLC according to the current task processed length and current machine speed, formula: remaining processing time = (total processing length - processed length) / current machine speed).

[0049] For the workshop topology and path information, exemplarily, collected from the workshop WMS (warehouse management system) and conveyor line PLC database, reflecting the physical connection and real-time load state of the workshop conveying network.

[0050] Specifically, it can include topology structure information, such as conveyor line connection relationship (e.g. the output end of “Conveyor-03” is connected to “Conveyor-05” and “Conveyor-07”, forming a passable path), path transmission delay empirical value (e.g. the average transmission time of “Conveyor-03→Conveyor-05” is 2 minutes, based on historical transmission data statistics).

[0051] Real-time load information, for example, path current load (unit: root, refers to the number of currently in-transit cones on a certain conveying line or a certain path, such as "Conveyor-03" currently in-transit cones 12 roots).

[0052] For historical scheduling records, exemplary can be collected from the MES scheduling log and the exception record module of the quality traceability system.

[0053] Specifically, it can include allocation relationship records, such as historical allocation (such as "TS-023" cone allocation to beaming machine one (ID "WJ-08")), and result labels (enumeration value: success / re-allocation / cause downtime / color difference problem, marked by MES according to task completion status and quality detection result).

[0054] Performance-related records, such as yield (actual output length of the beaming machine corresponding to this allocation, unit: meters), delay time (if the task is not completed on time, delay time = actual completion time - planned completion time, unit: minutes).

[0055] Preferably, according to the difference of data real-time requirement, the data with high real-time requirement such as beaming machine state and conveying line path load is set to a synchronization frequency of 1-10 seconds; the data with low real-time requirement such as cone task basic information and historical scheduling records is set to a synchronization frequency of 1-5 minutes. Through the intermediate database, the unified storage of multi-system data is realized, exemplary such as establishing an association index according to time stamp and equipment ID / cone ID, ensuring that the full amount of data of the same time dimension and the same object (such as a certain cone and a certain beaming machine) can be quickly called during subsequent preprocessing, and avoiding data fragmentation.

[0056] Preferably, the beaming workshop data is preprocessed, including setting a sliding time window to collect time series data, normalizing continuous numerical features, numerically encoding or embedding discrete classification features, and filling missing data.

[0057] Specifically, according to the principle of capturing short-term workshop dynamic changes and taking into account data timeliness and computational efficiency, exemplary the sliding time window length is set to 180-300 seconds. It can be understood that in actual application, the beaming workshop cone conveying period and the beaming machine task switching frequency can be flexibly adjusted (such as 180 seconds for short conveying period and 300 seconds for long conveying period), and the present application does not limit it.

[0058] The window sliding step is synchronized with the data collection frequency, and is exemplarily set to 5-10 seconds (matching the real-time sampling frequency of the PLC), ensuring that each sliding window covers the latest time series data without data redundancy. In an embodiment, a statistical summary is performed on the speed sequence in the window to calculate the average speed (reflecting the average conveying speed of the bobbin yarn in the period, avoiding interference from single-point abnormal values) and the speed slope (calculated by linear fitting, reflecting the speed change trend, such as a positive slope indicating an increasing speed and a negative slope indicating a decreasing speed), and the average speed and the slope are used as the time series feature summary of the subsequent model input, replacing the original long sequence data to reduce the computational complexity.

[0059] In an embodiment, the continuous numerical features are standardized, which, for example, filters out five types of continuous numerical features such as speed (including the average speed in the window), remaining bobbin length, path transmission delay, current path load (number of bobbins in transit), and estimated remaining processing time, and eliminates the dimensional influence through standardization. Exemplarily, z-score standardization or min-max standardization is adopted, and this embodiment will not be described again.

[0060] In an embodiment, the discrete classification features are numerically encoded or embedded, which, for example, for discrete features with explicit priority or hierarchical relationship, such as priority, tension level, and beaming machine task state, are numerically encoded.

[0061] Specifically, for priority encoding, it is, for example, mapped to a numerical value according to “urgency”, such as “urgent (rush order) = 3, regular = 2, low priority = 1”.

[0062] For tension level encoding, it is, for example, mapped to a numerical value according to “tension size”, such as “high tension = 5, medium-high tension = 4, medium tension = 3, medium-low tension = 2, low tension = 1”, and when matched with the “allowed tension range” of the beaming machine (continuous feature, already standardized), the tension difference can be quickly calculated (such as a bobbin with an encoding of 5 and a beaming machine with an allowed range of 3-4 has a tension conflict).

[0063] For beaming machine task state encoding, it is, for example, mapped to a numerical value according to “whether a new task can be accepted”, such as “idle = 1 (can be accepted immediately), tail yarn = 2 (will be idle soon), starting = 3 (ready to run, cannot be accepted temporarily), running = 4 (working), maintenance = 5 (cannot be accepted)”, and the encoded model can quickly filter out beaming machine nodes in the “cannot be accepted” state.

[0064] For discrete features without hierarchical relationship and only requiring categorical classification, such as color number (different color numbers have no priority difference, but need to be strictly matched with the target color number of the beaming machine to avoid color difference), embedding representation is performed.

[0065] Specifically, first, all color numbers are one-hot encoded. For example, if there are 10 color numbers in the workshop, color number A corresponds to the vector [1, 0, 0,..., 0], and color number B corresponds to [0, 1, 0,..., 0], ensuring the independence of different color number categories.

[0066] Next, low-dimensional embedding mapping is performed. Since the dimension of the one-hot vector is consistent with the number of color numbers (the more color numbers, the higher the dimension, which can easily lead to dimension disaster), it is converted into a 64-256 dimensional low-dimensional embedding vector (preferably matching the GNN node embedding dimension) through linear mapping.

[0067] Preferably, to address the short-term data shortage or long-term data loss that may occur during data collection, a hierarchical filling strategy is adopted to ensure data integrity and accuracy.

[0068] Specifically, for short-term missing data, such as temporary data interruption (e.g., sensor offline) or missing duration ≤1 sliding time window, the previous value / mean value filling method is used.

[0069] Specifically, the previous value filling method includes filling the value of the previous sampling point if the missing data is a time-series continuous feature (e.g., speed, path delay), such as filling the speed value at the 9th second to replace the missing speed data at the 10th second, to maintain the continuity of time-series data and avoid trend discontinuity.

[0070] The mean value filling method includes filling the mean value of the same batch / type data if the missing data is a non-time-series continuous feature (e.g., remaining tube length, path load) or an ordered discrete feature (e.g., tension level), such as calculating the mean value of the remaining tube length of other bobbins in the same batch (same dyeing batch) to fill the missing data of a certain bobbin, or taking the mode (highest frequency) of the tension level of the same type of warping machine to fill the missing data of a certain warping machine.

[0071] For long-term missing data, such as data missing duration >1 sliding time window, the historical mean value filling and weight adjustment are used to reduce the interference of missing data on the model.

[0072] Specifically, the historical mean filling includes extracting the historical data mean of the same type of object and the same process condition from the MES historical database as the filling value, such as the current machine of a certain warping machine extracting the historical data mean of the same type of object and the same process condition from the MES historical database as the filling value, and the uncertainty weight adjustment includes adding an uncertainty weight label (with a value of 0.5-0.8, and the normal data weight is 1.0) next to the filled data, and the weight size is negatively correlated with the missing time (the longer the missing time, the lower the weight). When calculating the node / edge features of the subsequent GNN model, the contribution of missing data will be reduced according to the weight (such as the filling machine speed with a weight of 0.5, which only accounts for 50% of the influence weight in speed matching degree calculation), avoiding false data guiding model decision.

[0073] In some embodiments, for S2, the cone yarn task node ( ), the screening range is the cone yarn task currently in the conveying line to be allocated and about to be hung (within a future 1 sliding time window, it can enter the allocation process), each node corresponds to an independent cone yarn, and the cone yarn number is used as the unique identifier to avoid repetition or omission.

[0074] The warping machine node ( ) screens the warping machine that can currently undertake new tasks and will be idle (the predicted remaining processing time is less than or equal to the sliding time window length), each node corresponds to the current task slot of one warping machine (each warping machine has only one task slot, i.e., one time at most one cone yarn task is undertaken), and the warping machine ID (such as “WJ-08”) is used as the unique identifier to exclude unavailable warping machines that are in maintenance, running, and have a long remaining processing time.

[0075] For directed edges, obviously incompatible pairs are first excluded to reduce subsequent calculation amount, and only cone yarn-warping machine pairs that meet the three basic conditions of color number matching, path reachability, and tension constraint are reserved to construct directed edges (the direction is from the cone yarn task node to the warping machine node, indicating that the cone yarn can be transported to the warping machine).

[0076] Specifically, for the color number matching constraint, it is compared with the target color number of the warping machine, and only the pairs that are completely consistent or process-allowed similar color numbers (such as some orders allowing light / dark color numbers of the same color system to be mixed) are reserved, and the pairs with large color number difference that are easy to produce color difference (such as red cone yarn and blue target color number of the warping machine) are directly excluded.

[0077] For path reachability constraint, it determines whether the current delivery line where the bobbin is located can reach the corresponding upper hanging delivery line of the warping machine through the existing delivery network based on the inter-plant topology and path information. For example, by querying the preset delivery line connectivity table, if there is at least one path without congestion and failure (the current load of the path is less than or equal to 90% of the rated capacity to avoid congestion leading to timeout), it is determined to be reachable and the pairing is retained. Otherwise (such as no connected path or full congestion of the path), it is excluded.

[0078] For tension constraint, it compares the bobbin tension level with the allowable tension range of the warping machine, and only retains the pairing where the bobbin tension level falls within the allowable tension range of the warping machine or the tension difference is less than or equal to 1 level (such as the allowable tension range of the warping machine is 3-4 levels, and the bobbin tension level is 2 or 5 levels, which is a tension difference of 1 level, and the process allows short-term adaptation), and excludes the pairing where the tension difference is greater than or equal to 2 levels, which is likely to cause yarn breakage or uneven winding (such as a 5-level high-tension bobbin and a 2-level low-tension warping machine without corresponding edges).

[0079] Preferably, the physical properties, process requirements, and real-time state of the bobbin task are converted into a numerical feature vector, and the dimension is adapted to the GNN embedding dimension (such as 64-256 dimensions, and the feature items are expanded to the corresponding dimension through linear mapping). Specifically, the bobbin task node includes a standardized speed sequence digest, a remaining bobbin length, a color number embedding vector, a tension level, a priority, and a current path load feature.

[0080] Specifically, for the standardized speed sequence digest, it uses the average speed (which has been standardized by z-score) and the speed slope (calculated by linear fitting, such as the speed increasing from 20 m / min to 22 m / min in the last 300 seconds, the slope is 0.0067 m / (min·s), and the value after standardization) within the sliding time window to reflect the average level and trend of the bobbin delivery speed.

[0081] For the remaining bobbin length, it is normalized to the [0, 1] interval (such as a full bobbin length of 1000 m, and a remaining bobbin length of 500 m, which is standardized to 0.5), which represents the processable amount of the bobbin.

[0082] For the color number embedding vector, it is a low-dimensional embedding vector (64-256 dimensions) mapped after one-hot encoding of the bobbin color number, which retains the color number category information and avoids dimension disaster.

[0083] For the tension level, it is a numerically encoded tension level value (such as low tension = 1, medium tension = 3, and high tension = 5), which is convenient for comparison with the allowable tension range of the warping machine.

[0084] For the priority, it is a numerically encoded priority value (low priority = 1, regular = 2, and urgent = 3), which reflects the task urgency.

[0085] For current path load, its, for example, the numerical value of the real-time load of the conveying line where the current cone yarn is located (the number of cones in transit) after min-max standardization (such as the rated capacity of the conveying line is 20, and the current is 10, which is standardized to 0.5), reflecting the congestion risk of the current conveying environment of the cone yarn.

[0086] Preferably, the running state, process requirements, and scheduling requirements of the beaming machine are also converted into numerical feature vectors. Specifically, the beaming machine node includes the current machine speed, target color number embedding vector, allowable tension range, state code, priority, and predicted idle time features.

[0087] Specifically, for the current machine speed, its, for example, the numerical value of the real-time speed of the beaming machine spindle after z-score standardization (based on the historical machine speed mean and standard deviation), reflecting the current processing speed level.

[0088] For the target color number embedding vector, its, for example, a low-dimensional vector homologous to the cone yarn color number embedding vector (ensuring consistent color number feature space), facilitating the calculation of color number compatibility.

[0089] For the allowable tension range, its, for example, the "minimum" and "maximum" of the allowable tension range are encoded respectively (such as allowing 2-4 levels, which is encoded as [2, 4], and then standardized to expand to the embedding dimension), retaining the tension adaptation interval information.

[0090] For the state code, its, for example, the beaming machine task state after numerical encoding (idle = 1, tail yarn = 2, start = 3, run = 4, maintenance = 5), directly identifying whether the beaming machine can take on a new task.

[0091] For the priority, its, for example, a numerical value consistent with the coding rules of the cone yarn task priority (such as the beaming machine priority associated with an urgent order = 3), ensuring that the priority dimension can be compared across nodes.

[0092] For the predicted idle time, its, for example, the time from the completion of the current task of the beaming machine to the time when it can take on a new task (such as 10 minutes of remaining processing time plus 2 minutes of downtime for refueling, with a predicted idle time of 12 minutes), which is standardized to the [0, 1] interval, reflecting the time availability of the beaming machine.

[0093] Preferably, for the adaptability, historical performance, and transmission cost of the cone-beaming machine pair, an edge feature vector is constructed to quantify the pros and cons of their association. Specifically, the directed edge includes speed matching degree, path transmission delay estimation, historical pairing success rate, color number compatibility score, and tension penalty value features.

[0094] Specifically, for the speed matching degree, it is calculated by, for example, the standardization speed average of the bobbin and the standardization current speed of the beamer (if the ratio is within the range of 0.9-1.1, the matching degree = 1.0; the greater the deviation of the ratio, the matching degree decreases linearly, such as 0.8, the matching degree = 0.8), reflecting the speed coordination (small speed difference can reduce the risk of yarn breakage).

[0095] For the path transmission delay estimation, it is calculated, for example, based on the workshop topology and path information, the shortest path transmission time of the bobbin from the current position to the beamer (such as 2 conveying lines, the total delay = the sum of the path delay experience values of each section), and is normalized to the interval [0, 1] by min-max, the shorter the delay, the higher the score.

[0096] For the historical pairing success rate, it is calculated, for example, by querying the historical scheduling records, and the proportion of the number of successful assignments of the type of bobbin (same batch, same color number, same tension level) to the beamer to the total number of assignments (such as 9 successful assignments out of 10 historical assignments, the success rate = 0.9), directly reflecting the historical reliability of the pairing.

[0097] For the color number compatibility score, if the color number of the bobbin is completely consistent with the target color number of the beamer, the score = 1.0; if it is a similar color number allowed by the process, the score = 0.8-0.9 (adjusted according to the color difference value, the smaller the color difference value, the higher the score); and no compatible relationship, the score = 0 (already excluded in the edge screening stage), quantifying the risk of color number adaptation.

[0098] For the tension penalty value, if the tension level of the bobbin falls within the allowable tension range of the beamer, the penalty value = 0; if the tension difference = 1 level, the penalty value = 0.1 (slight penalty); and the tension difference ≥ 2 levels (already excluded in the edge screening stage, this is a redundant check), the penalty value = 1.0 (severe penalty), avoiding tension conflicts through the penalty mechanism.

[0099] In some optional embodiments, to solve the problem that the static graph cannot reflect the dynamic changes of the bobbin position, the beamer state update, and the path load fluctuation, a time sequence dynamic graph is constructed. Specifically, the bipartite graph model is a time sequence dynamic graph, the time sequence change characteristics of the bobbin task and the beamer are integrated into the feature representation of the node and the edge by introducing time coding or using a sliding time window, so as to capture the dynamic evolution of the workshop state.

[0100] Specifically, in an embodiment, the sliding time window length (such as 180-300 seconds) is taken as a period, and the vertex screening-edge screening-feature vector generation process is periodically re-executed to realize dynamic updating of the graph structure. Illustratively, for vertex updating, the bobbin nodes that have completed assignment and the beamer nodes that enter the maintenance state are removed; the bobbin nodes that have just entered the conveying line and are to be assigned and the beamer nodes that are about to be idle are added, so as to ensure that the vertex always covers the current active scheduling object.

[0101] For edge update, based on the latest color number demand, path load, tension state, re-screen the directed edge (such as a path from congestion to smooth, add the corresponding edge; the tension range of the warping machine is adjusted, and the edge that does not meet the requirements is deleted), to avoid invalid edges based on outdated information.

[0102] For feature vector update, for example, the eigenvalues of each node and edge are recalculated (such as the remaining bobbin length of the bobbin yarn is reduced, the expected idle time of the warping machine is shortened, and the path transmission time delay changes), to ensure that the feature vector can reflect the object state in real time.

[0103] In one embodiment, time coding is added to the node and edge feature vectors. Exemplary specific implementation includes relative timestamp coding, taking the current scheduling decision time as the reference, calculating the relative time of the bobbin yarn arriving at the current conveying line (such as arriving 10 minutes ago, relative time =-600 seconds), the relative time of the expected idle time of the warping machine (such as idle 12 minutes later, relative time = 720 seconds), mapping the relative time into a periodic vector (such as sin(2πt / T), cos(2πt / T), T is the time period, such as 3600 seconds) through the sin / cos function, and embedding the node feature vector.

[0104] Thus, through the above timing expansion method, the bipartite graph can be updated in real time to follow the changes in the workshop state, ensuring that the subsequent GNN model can output matching scores that fit the actual situation based on dynamic graph structures, avoiding the problem of decision lag caused by static graphs.

[0105] In some embodiments, for S3, as shown in Figure 2 The architecture of the GNN model includes a node embedding layer, a multi-message passing layer, and an edge scoring head, wherein the node embedding layer is used for uniform feature dimension and normalization, the multi-message passing layer is used for bidirectional information flow and attention aggregation (exemplarily, 2-4 layers), and the edge scoring head is used for outputting comprehensive matching scores.

[0106] Specifically, in one embodiment, the node embedding layer maps the bobbin task node feature vector and the warping machine node feature vector to the same dimension hidden state space (for example, embedding dimension 64-256 dimensions, which can be flexibly adjusted according to the workshop data size and computing power), eliminating the difference in original feature dimensions.

[0107] For uniform feature dimension, for example, the 6-dimensional original features of the bobbin node (standardized speed sequence summary, remaining bobbin length, etc.) and the 6-dimensional original features of the warping machine node (current machine speed, target color number embedding, etc.) are respectively expanded to the preset embedding dimension through the linear mapping layer (such as 6-dimensional to 64-dimensional), and the dispersed feature information is integrated into a high-dimensional vector through the learnable weight matrix during the mapping process.

[0108] For normalization, for example, LayerNorm normalization is performed on the linear mapped node feature vector, which can be simplified as the formula: ; wherein, is the feature mean, is the feature variance, is a learnable parameter, is a small value to prevent the denominator from being 0), to avoid the dominance of some feature values in model learning (such as the remaining tube length standardized to 0-1, and the color number embedding vector to -1-1, which can be in the same order of magnitude after normalization).

[0109] After the above processing, the initial hidden state of the tube yarn node is output and the initial hidden state of the beaming machine node is , both of which have the same dimension (such as 64 dimensions) and can directly participate in subsequent message passing.

[0110] In one embodiment, the message passing layer is the core of the GNN to capture the association between nodes. In one embodiment, a bidirectional message passing mechanism (tube yarn→beaming machine, beaming machine→tube yarn) is used, and the neighbor messages are aggregated by attention mechanism weighting. Each layer is updated based on the node hidden state of the previous layer.

[0111] Specifically, as shown in Figure 3 , the bidirectional message passing process includes the following steps: S301, message passing layer parameter setting; wherein, the number of layers is selected, for example, according to the workshop node size (such as 50-200 tube yarn nodes and 10-50 beaming machine nodes), it is set to 2-4 layers, too few layers (<2 layers) cannot fully fuse cross-node information, too many layers (>4 layers) are easy to cause gradient disappearance and calculation delay, and 2-4 layers can balance the feature fusion effect and engineering practicability.

[0112] The type of attention mechanism, for example, uses graph attention, which allows the model to automatically learn the importance weight of different neighbor nodes, avoiding the masking of key node information by traditional sum / mean aggregation (such as the influence of high-priority tube yarn on the beaming machine should be greater than that of low-priority tube yarn).

[0113] S302, forward message passing; tube yarn node→beaming machine node, beaming machine learning tube yarn feature; specifically, taking the lth layer→l+1th layer as an example, the first step: generating a message vector from the tube yarn to the beaming machine For each directed edge from a tube yarn node S to a beaming machine node M, based on the feature vector (speed matching degree, historical success rate, etc.), the lth layer hidden state of the tube yarn node , the lth layer hidden state of the beaming machine node , the message function generates a message vector: ; where "||" represents feature concatenation (e.g., 64-dim + 64-dim + 64-dim = 192-dim), is a linear mapping function (192-dim → 64-dim), which compresses the concatenated high-dimensional features to the embedding dimension while incorporating the associated information of edge features (e.g., higher speed matching degree results in larger corresponding weight in the message vector).

[0114] Second step: Attention weight calculation calculates the attention weight of each neighbor for all neighbor beam nodes (i.e., beam connected to M) of the warping machine node M. , which can be simplified as: ; where is the attention weight matrix (128-dim → 1-dim), is the activation function, is the dimension normalization factor (d is the embedding dimension), ensures that the sum of all neighbor attention weights of the warping machine M is 1. High-priority, high-speed matching degree beam nodes will get higher (e.g., 0.8), and low-priority beams have lower weights (e.g., 0.1).

[0115] Third step: Warping machine node hidden state update sums all neighbor message vectors of the warping machine node M according to the attention weight to get the aggregated message , and then fuses it with its own l-th layer hidden state to update it to the (l+1)-th layer hidden state , which can be simplified as: ; where is the linear transformation matrix, is the bias term, The activation function introduces nonlinearity, allowing the warping machine node features to contain both its own state and the key information of neighbor beams (e.g., the warping machine can perceive the pending allocation demand of high-priority beams).

[0116] S303, backward message passing, warping machine node → beam node, beam learns warping machine features, process is symmetric with forward passing, allowing beam nodes to perceive the state constraints of warping machines (e.g., whether the warping machine is about to be idle, whether the allowable tension range is suitable), avoiding excessive assignment of beams to a warping machine.

[0117] Specifically, generate the message vector from the warping machine to the beam: ; where the edge feature is reused to ensure consistency of associated information.

[0118] Calculate the attention weight of the beam node to the neighbor warping machine: .

[0119] The weight of the idle warper is higher than that of the running warper; the cone yarn node hidden state is updated: ; wherein the meanings of the parameters correspond to the formula parameters in step S202, and will not be described here.

[0120] S304, multi-layer iteration update, repeat the above bidirectional message passing process 2-4 layers, each layer is based on the hidden state optimization feature expression of the last layer, wherein, with the increase of the number of layers, the node hidden state gradually fuses the multi-order association information of itself-direct neighbor-indirect neighbor (such as the cone yarn can not only perceive the directly connected warper, but also perceive the load condition of other cone yarns adjacent to the warper through the warper), and finally outputs the node hidden state of the Lth (such as the 4th layer) (cone yarn task) and (warper).

[0121] In one embodiment, the core of the edge score head is to calculate the comprehensive matching score score(S, M) of each directed edge S→M based on the node hidden state of the final layer and the edge feature, and the score range is mapped to [0, 1] (1 represents the optimal matching, and 0 represents the worst matching), and exemplarily, the specific implementation can include: feature splicing, including splicing the final hidden state of the cone yarn node , the final hidden state of the warper node, and the edge feature vector into a high-dimensional feature vector, such as 64+64+64=192.

[0122] Linear transformation and activation, including compressing and nonlinearly transforming the spliced features through a 2-layer fully connected network, exemplarily, the first layer maps the 192-dimensional features to 32-dimensional features, and the ReLU activation is performed; the second layer maps the 32-dimensional features to 1-dimensional features, and finally the output value is constrained in the [0, 1] interval through the Sigmoid activation function, and the final matching score is obtained.

[0123] Wherein, the score high and low comprehensively reflects the speed matching degree, color number compatibility, tension adaptability, historical success rate, priority matching and other multi-dimensional indicators (such as a score of 0.9 indicates that the pair is optimal in quality, efficiency and priority, and a score of 0.2 indicates that there is a significant constraint conflict).

[0124] In an optional embodiment, the training of the graph neural network model adopts a multi-objective joint loss function for supervised training; the joint loss function at least includes: a supervised learning loss based on historical successful pairing data, a ranking comparison loss for ensuring that high-quality pairing scores are higher than low-quality pairing, and a constraint-aware loss for imposing a penalty on pairing that violates the tension or path capacity constraint.

[0125] Specifically, first, the training data is prepared, including constructing edge-level training samples, specifically, according to the time sequence, the historical scheduling time points (such as every 5 minutes in the past 6 months) are backtracked, and for each time point, the corresponding bipartite graph (including node features, edge features, and edge existence) is reconstructed to simulate the workshop state at that time.

[0126] By sampling positive samples / under-sampling negative samples, the ratio of positive and negative samples is ensured to be close to 1:1, avoiding model bias towards the majority class; wherein the positive sample is an edge that is actually allocated in history and the result label is success, for example, S1→M1 is not broken and has no color difference after actual allocation, and the label y=1 is marked, and each positive sample corresponds to one high-score edge; wherein the negative sample is an edge that is not allocated in history or the result label is re-allocated / stop / mismatch after allocation, for example, S2→M1 is not allocated and S3→M2 is allocated after stop, and the label y=0 is marked, and each negative sample corresponds to one low-score edge.

[0127] In one embodiment, a hybrid loss function of supervision, comparison and constraint is adopted, which is simplified as follows: ; is a weight coefficient, and the sum is 1, such as 0.3, 0.4, 0.2, 0.1, and each loss term is implemented as follows: supervised learning loss , fitting historical successful pairing; specifically, based on the label of historical positive / negative samples, cross-entropy loss is used to measure the difference between the model prediction score and the true label, which can be simplified as follows: ; wherein N is the number of edge instances in the batch, is the prediction score of the i-th edge by the model. When the model prediction score of the positive sample is close to 1 and the negative sample is close to 0, the loss value is the smallest, ensuring that the model learns the feature rules of historical successful pairing (such as edges with high speed matching degree and high historical success rate are easy to be predicted as high score).

[0128] Ranking comparison loss , ensuring the reasonableness of the score ranking, specifically, for multiple candidate beams of the same warping machine or multiple candidate warping machines of the same beam, pair wise hinge loss is used to force the high-quality pairing score to be higher than the poor-quality pairing, which can be simplified as follows: ; wherein K is the number of high-quality-poor-quality pairings (P) in the batch, is the high-quality pairing score (such as the successfully allocated edge), is the poor-quality pairing score (such as the unallocated edge), and margin is a preset threshold (such as 0.2, ensuring that the high-quality and poor-quality scores have a difference of at least 0.2). When , the loss is 0; otherwise, the loss increases as the difference decreases, avoiding the unreasonable situation that the model appears the high-quality pairing score lower than the poor-quality pairing (such as the successfully allocated edge score 0.6 and the unallocated edge score 0.7).

[0129] Constraint-aware loss , to circumvent hard constraint conflicts, specifically, for the pairings that violate tension range or exceed path capacity, a penalty term is introduced to force the model to learn executable solution space, simplified as formula: ; where is an indicator function (1 if the ith edge violates the constraint, otherwise 0), is a penalty coefficient (to ensure the penalty is strong enough), and penalty is a fixed penalty value (e.g., 1.0). For example, if the tension level of an edge exceeds the range allowed by the beamer (violates the constraint), the constraint loss of this edge is , which significantly increases the total loss, prompting the model to actively reduce the score of such illegal pairings (e.g., from 0.5 to 0.1).

[0130] Regularization loss , to prevent the model from overfitting, specifically, L2 regularization is used to constrain the model parameters (e.g., weight matrices of linear mapping layers and attention layers), formula: ; where W is the set of all learnable parameters of the model, is the L2 norm of the parameters. This loss term increases with the absolute value of the parameters, avoiding the model's excessive reliance on partial features (e.g., only focusing on priority while ignoring tension constraints), and improving the model's generalization ability (e.g., still reasonably scoring when facing new color yarns).

[0131] In some embodiments, for S4, obviously incompatible pairings are removed in advance to reduce the computational load of subsequent global matching, around the preset color compatibility, tension range, and path reachability hard constraints, a hierarchical filtering rule and verification mechanism is established to ensure that the remaining directed edges meet the basic process and physical limitations of the workshop production.

[0132] Specifically, before performing the filtering, the baseline parameters required for constraint judgment are extracted to ensure that the rules are quantifiable and executable, and all parameters are homologous to the coarse filtering parameters of S2, but the judgment accuracy is higher (S2 is "coarse filtering" to reduce the initial number of edges, and S4 is "fine filtering" to exclude potentially illegal edges). Exemplarily, the specific parameters sources and settings include: color compatibility baseline: read the color compatibility list of the current production order from the MES process parameters, clearly define "completely compatible color" (e.g., order requires color #FF0033, only #FF0033 is completely compatible) and "technically allowed approximately compatible color" (e.g., some blended orders allow #FF0033 and #FF0044 (light degree of the same color system) to be compatible, and need to mark the color difference value ≤ΔE2.0, ΔE is the international standard color difference value), color combinations that do not meet the list are directly judged as incompatible.

[0133] Tension range reference: read the “real-time allowed tension range” of each warper from the PLC control system of the warper (different from the “static allowed range” of S2, here is the dynamic range considering the current running state of the machine, such as the tension accuracy of the warper decreases slightly after running for 1 hour, the allowed range is adjusted from 20-25N to 19-26N), and extract the “actual tension detection value” of each cone from the cone task table (not the “tension level” of S2, it is the real-time value collected by the laser tension meter, unit N), to ensure that the tension comparison is based on accurate values rather than levels.

[0134] Path reachability reference: obtain the “current load rate” of each conveying line (in-transit cone number / capacity, such as the conveying line Conveyor-03 has a rated capacity of 20 cones, and currently has 18 in-transit, the load rate is 90%), “device running state” (normal / fault / maintenance), “transmission delay threshold” (such as the maximum allowed transmission delay from the current conveying line to the warper hanging point, set according to the expected idle time of the warper, such as the warper is idle after 10 minutes, the transmission delay threshold is set to 8 minutes, reserving 2 minutes for hanging preparation), paths that do not meet the physical transmission restrictions are determined as unreachable.

[0135] Based on the above reference parameters, three major hard constraint filters are executed in layers. Specifically, the first layer: color number compatibility filtering, excluding color difference risk edges. Exemplarily, a dual judgment logic of accurate matching and color difference value verification is adopted to remove directed edges that do not meet the compatibility requirements. The specific steps are as follows: First step: complete compatibility color number matching: compare the “actual color number” of the cone node with the “target color number” of the warper node, if they are completely consistent (such as the color number of the cone is #FF0033 and the target color number of the warper is #FF0033), it is directly determined as compatible and the edge is retained; if they are not consistent, proceed to the second step of approximate compatibility verification.

[0136] Second step: approximate compatibility color difference value verification: if the color number of the cone and the warper is not consistent, query the “color number compatibility list” to confirm whether it belongs to the “process allowed approximate compatibility” range.

[0137] If not in the list, directly remove the edge (such as the cone #FF0033 and the warper #00FF00 are not in the list, it is determined as incompatible).

[0138] If it is in the list, extract the “historical color difference value record” of the cone and the target color number of the warper (such as the average ΔE1.5 of the same color number combination in the past 3 months), and collect the “real-time color difference value” of the cone by the workshop color difference instrument (such as ΔE1.8), if the real-time color difference value ≤ the threshold value marked in the list (such as ΔE2.0), it is determined as compatible and the edge is retained; if the real-time color difference value > the threshold value (such as ΔE2.5), it is determined as incompatible and the edge is removed.

[0139] Optionally, for multi-cone warping tasks (multiple cones need to be hung at the same time), additional checks are performed on the color difference values between all candidate cones on the same warping machine, and the color difference value between the cones is required to be ≤ΔE1.0 to avoid uneven appearance of the finished product caused by color difference between the cones. If the color difference value of any two cones exceeds the standard, all related edges corresponding to the warping machine are removed.

[0140] Second layer: tension range filtering, excluding broken yarn / winding uneven risk edges, exemplarily, ensuring stable tension during warping process, avoiding broken yarn or uneven winding density, based on accurate tension value rather than grade, establishing numerical interval matching and dynamic margin checking filtering logic, removing directed edges with excessive tension, the specific steps are as follows: first step: tension value interval matching: extract the actual tension detection value of the cone (such as 22N) and the real-time allowed tension range of the warping machine (such as 19-26N), if the actual tension value of the cone falls within the allowed tension range of the warping machine, it is directly determined as tension compatible, and the edge is retained; if it exceeds the range (such as cone tension 28N, exceeding the upper limit of 26N), enter the second step of dynamic margin checking.

[0141] Second step: dynamic margin checking: considering that there is a small fluctuation (such as ±1N) in the tension during the operation of the warping machine, a "dynamic margin threshold" is set (usually 5% of the upper limit of the allowed range, such as 26N upper limit of the allowed range, dynamic margin 1.3N, actual acceptable upper limit 27.3N).

[0142] If the actual tension value of the cone ≤ the upper limit of the allowed range + the dynamic margin (such as 28N>27.3N, exceeding the margin), or ≥ the lower limit of the allowed range - the dynamic margin (such as cone tension 17N<19-1.3=17.7N, exceeding the margin), it is determined that the tension is not compatible, and the edge is removed.

[0143] If the actual tension value of the cone is within the "allowed range ± dynamic margin" (such as cone tension 27N≤27.3N), and the warping machine PLC feedback "current tension control accuracy meets the standard" (such as tension fluctuation ≤±0.5N), it is determined to be compatible, and the edge is retained; if the warping machine tension control accuracy does not meet the standard (fluctuation >±1N), even if the value is within the margin, the edge is still removed (to avoid fluctuation leading to temporary exceeding of the range).

[0144] Third layer: path reachability filtering, excluding delivery timeout / congestion risk edge, ensuring that the cheese can reach the upper hanging point before the warper is idle, avoiding the congestion of the conveying line, based on the real-time path state, establishing a triple filtering logic of path connectivity, load rate and time delay, eliminating the directed edge that is not reachable or transmission timeout, the specific steps are as follows: First step: path connectivity verification: through the workshop topology and path information, query the "shortest path" (preferably, based on Dijkstra algorithm calculation, prefer to select the path with the shortest transmission time delay) from the current conveying line (such as Conveyor-01) where the cheese is located to the corresponding upper hanging conveying line (such as Conveyor-08) of the warper, if there is any section of conveying line in "failure" or "maintenance" state (such as Conveyor-05 failure in the path), and there is no other alternative path (such as no Conveyor-01→Conveyor-06→Conveyor-08 backup path), it is judged that the path is not connected, and the edge is removed; if there is at least one normal path, enter the second step of load rate verification.

[0145] Second step: path load rate verification: calculate the "average load rate" of all conveying lines in the shortest path (such as the path containing Conveyor-01 (load rate 70%), Conveyor-06 (load rate 85%), average load rate 77.5%), and check whether there is a "congestion conveying line with load rate ≥95%" (such as Conveyor-07 with load rate 98%, judged as congestion).

[0146] If the average load rate is >90% or there is a congested conveying line, it is judged that the path load is too high, which is easy to cause the increase of transmission time delay even if it is connected, and the edge is removed.

[0147] If the average load rate is ≤90% and there is no congested conveying line, enter the third step of transmission time delay verification.

[0148] Third step: transmission time delay verification: based on the experience value of path transmission time delay (such as the experience time delay of Conveyor-01→Conveyor-06 is 2 minutes, the experience time delay of Conveyor-06→Conveyor-08 is 3 minutes, the total experience time delay is 5 minutes), the time delay is corrected combined with the current load rate (the load rate increases by 10%, the time delay increases by 10%, such as the average load rate is 80%, the time delay is corrected to 5 minutes x (1+0%) = 5 minutes; the load rate is 90%, the correction is 5 minutes x (1+10%) = 5.5 minutes), and the "corrected transmission time delay" is obtained.

[0149] If the corrected transmission time delay is ≤ the "transmission time delay threshold value" of the warper (such as the warper is idle after 10 minutes, the threshold value is 8 minutes, 5.5 minutes ≤ 8 minutes), it is judged that the path is reachable, and the edge is retained.

[0150] If the modified transmission delay > threshold (e.g., the modified delay is 9 minutes > 8 minutes), it is determined that the transmission is timed out (the beam reaches the warping machine which has started the next task), and the edge is rejected.

[0151] Thus, through the above process, S4 can effectively reject incompatible edges, greatly reducing the computational complexity of the global matching in S5, while ensuring that all retained edges meet the basic production constraints, laying the foundation for generating a feasible scheduling scheme.

[0152] In some embodiments, for S5, based on the bipartite graph filtered by S4, the beam-warping machine assignment is converted into a mathematically solvable weighted bipartite graph matching problem, ; wherein the vertex set: the set of beam task nodes defined in S2 (filtered beams to be assigned, the number is n), and the set of warping machine nodes (filtered warping machines that can take on tasks, the number is m), ensuring that the vertices only contain valid scheduling objects.

[0153] Edge set: only keep the directed edges E' filtered by S4 (edges that meet color number, tension, and path constraints), each edge corresponds to a candidate scheme for assigning beam s to warping machine m.

[0154] Weight: the weight of each edge (s, m) is the matching score score(s, m) output by the GNN model in S3 (value 0-1), the higher the score, the better the comprehensive adaptability (quality, efficiency, priority) of the pair, which is the basis for maximizing the total matching score.

[0155] The constraint conditions convert the actual production restrictions into mathematical constraints to avoid invalid schemes such as "one beam to multiple machines" and "one machine to exceed capacity", and the specific constraints include: beam assignment constraints, for example, each beam task is assigned to at most one warping machine, i.e. for any beam s, at most one edge in all associated edges (s, m) is selected (the assignment variable indicates successful assignment, indicates no assignment), and the mathematical expression is: .

[0156] Warping machine capacity constraints, for example, the number of tasks received by each warping machine does not exceed its task slot capacity, and the warping machine is usually single-task slot (only one beam task can be accepted at a time), i.e. for any warping machine m, at most one edge in all associated edges (s, m) is selected, and the mathematical expression is: .

[0157] The objective function is constructed to maximize the global total matching score: ; that is, by selecting the appropriate (0 or 1) maximizes the sum of the weights of all selected edges, ensuring that the final allocation scheme achieves optimal fit overall, rather than local optimum (e.g., avoiding selecting high-scoring yarn for a warping machine, which would result in multiple yarns not being matched with warping machines).

[0158] Choose an appropriate solution based on the scale of the yarn package-warping machine (i.e., the size of n and m), balancing "solution accuracy" and "computational efficiency." Specifically, for small-scale scenarios ( and For example, when the number of yarn packages to be allocated and the number of warping machines that can be accepted are small (such as local scheduling in the workshop or emergency order replenishment scenarios), the Hungarian algorithm is used to solve the global optimum. Specifically, (1) Algorithm preprocessing: Constructing the cost matrix: Since the standard Hungarian algorithm is usually used for "minimum cost matching", it is necessary to transform "maximizing the total score" into "minimizing the total loss": Calculate the maximum score threshold max_score (take the maximum value of all edge score(s,m), such as 0.95); Construct The loss matrix C, matrix elements (If (s,m) is not filtered afterward, Set it to max_score+1 to indicate that this pairing is not optional. At this point, "minimize the total loss" "This is equivalent to "maximizing the total score Z".

[0159] (2) Execute the Hungarian algorithm: Step 1, matrix row reduction: Subtract the minimum value of each row of the loss matrix so that each row has at least one 0 element (representing the relatively best candidate pair in the row).

[0160] Step 2, Matrix column reduction: Subtract the minimum value of each column of the row-reduced matrix to ensure that each column has at least one 0 element, further simplifying the matching selection.

[0161] Step 3, Cover 0 elements: Cover all 0 elements with the fewest horizontal and vertical lines. If the number of covering lines is equal to min(n,m) (the smaller value between the number of yarn packages and the number of warping machines), a match can be found directly; otherwise, proceed to step 4.

[0162] Step 4, Adjust the matrix: Calculate the minimum value of the uncovered elements, subtract the minimum value from the uncovered rows, and add the minimum value to the covered columns. Repeat step 3 until the number of covering lines equals min(n,m).

[0163] Step 5, Find the optimal match: Select a pair from the position containing 0 elements, ensuring that each yarn package is selected only once by the warping machine. This minimizes the total loss, i.e., maximizes the total score. Output. The set of pairs.

[0164] The Hungarian algorithm has a time complexity of O(k^3 (k=min(n,m))). The time consumption can be controlled within 1 second (meeting the real-time scheduling requirement of the workshop), and the global optimal solution can be ensured, which is suitable for small-scale and high-precision scenarios (such as high-priority order scheduling).

[0165] For large-scale scenarios (k > 50) Or When the number of to-be-allocated packages or the number of available beaming machines is large (such as batch scheduling in the whole period of the workshop), the computational complexity of the Hungarian algorithm increases sharply (for example, the complexity reaches 125,000 when k = 50). The present embodiment uses a minimum cost flow algorithm based on maximum flow to improve the computational efficiency while ensuring the "approximate optimality". The specific steps include: (1) constructing a flow network model: converting the bipartite graph into a flow network of "source point - package - beaming machine - sink point", realizing the mapping of "cost" and "flow", and specifically, adding nodes: adding a source point S (the starting point of the flow) and a sink point T (the ending point of the flow), connecting all package nodes to the source point, and connecting all beaming machine nodes to the sink point, setting the edges, capacity and cost: the edge from the source point S to the package s: the capacity is set to 1 (each package can flow out at most 1 unit of flow, corresponding to 1 allocation), and the cost is set to 0 (no additional cost); the edge from the package s to the beaming machine m (only the filtered edge is reserved): the capacity is set to 1 (each pair can flow at most 1 unit of flow), and the cost is set to (the score is converted into cost in accordance with the Hungarian algorithm); the edge from the beaming machine m to the sink point T: the capacity is set to 1 (each beaming machine can flow in at most 1 unit of flow, corresponding to 1 acceptance), and the cost is set to 0.

[0166] (2) Perform minimum cost maximum flow solution: the goal is to find the maximum flow from S to T (i.e., the maximum number of package-beaming machine pairs that can be allocated), and the total cost is minimized under the maximum flow (i.e., the total score is maximized), which specifically uses the successive shortest augmenting path algorithm (successive shortest augmenting path algorithm), and the specific steps include: step 1, initialization: set the flow of all edges in the flow network to 0, the total cost to 0, and the maximum flow to 0.

[0167] Step 2, find the shortest path: use the Bellman-Ford algorithm or the SPFA algorithm (suitable for scenarios with negative cost edges) to find the minimum cost path from S to T (the sum of the costs of all edges on the path is the minimum).

[0168] Step 3, augment the flow: in the found shortest path, calculate the maximum flow that can be augmented (here, the capacity of all edges is 1, so the augmented flow is 1), add 1 to the flow of each edge on the path, add the path cost to the total cost, and add 1 to the maximum flow.

[0169] Step 4, repeat iteration: repeat steps 2-3 until no augmented path from S to T can be found (i.e., the maximum flow is reached), at which time the total cost is minimized, and the corresponding flow paths are the optimal match (the path from the source point to the bobbin to the beamer to the sink point corresponds ).

[0170] The time complexity of this algorithm is (F is the maximum flow, E is the number of edges, and V is the number of nodes), when n=100, m=50, the maximum flow is usually ≤50, the calculation time can be controlled within 3 seconds (satisfying the real-time requirement), and it can handle large-scale nodes, suitable for full-scale scheduling scenarios in workshops; although it cannot guarantee absolute global optimality in theory, the score difference with the global optimal solution is usually ≤5% in practical applications, and the accuracy can meet the production requirements.

[0171] From the algorithm output, all pairs of are extracted to form a "bobbin-beamer distribution list", which includes, for example: bobbin number, beamer ID, matching score, corresponding transport path (the shortest path of this pair is extracted from the S2 workshop topology and path information), predicted transmission delay (the corrected delay extracted from the S4 path reachability check), and predicted hanging time (current time and predicted transmission delay).

[0172] In some embodiments, for S6, each pair of "bobbin-beamer" distribution results is decomposed into transport line control instructions, beamer preparation instructions, and hanging coordination instructions to ensure that the entire process of moving the bobbin from the current location to the beamer for processing is covered, and the instruction format is consistent with the original production control of the production line, which will not be described here.

[0173] Preferably, in one embodiment, to timely detect deviations and provide early warning of abnormalities, two types of monitoring indicators are set, which are compared with the predicted values of the GNN model in S3 in real time.

[0174] Specifically, the key timing indicators are: monitoring the "actual transmission delay of the bobbin" (actual time - predicted time of the bobbin leaving the current transport line), and the "actual idle time of the beamer" (actual time - predicted time of the beamer completing the current task), and the comparison threshold is set to 20%, i.e., if the deviation between the actual value and the predicted value exceeds 20%, a warning is triggered (e.g., if the predicted transmission delay is 5 minutes and the actual transmission delay is 6.1 minutes, the deviation is 22%).

[0175] The key state indicators are: monitoring the "actual tension value of the bobbin" (compared with the tension adaptation value predicted by the GNN), the "actual load rate of the transport line" (compared with the predicted path load rate), and the "actual machine speed of the beamer" (compared with the predicted processing machine speed), and the comparison threshold is set to 10% (e.g., if the predicted tension is 22N and the actual tension is 24.5N, the deviation is 11.4%).

[0176] When any indicator deviates beyond the threshold, or a device failure signal is received (such as a conveyor line shutdown, a warping machine tension out of control), an abnormal response process is triggered immediately, the first step is to determine the re-planning range (local / global), including, the local re-planning determination condition: the abnormality only affects "a single cone yarn" or "a local device cluster" (such as the speed of a certain cone yarn is abnormal during the transmission of Conveyor-05, only affecting the cone yarn and the corresponding warping machine one (ID for WJ-08); or Conveyor-05 failure, only affecting 3 cone yarns that need to pass through the conveying line), and does not affect the normal execution of other conveying lines and warping machines, it is determined as "local abnormality", and local re-planning is started.

[0177] The global re-planning determination condition is: the abnormality affects "global resources" (such as a power fluctuation in the workshop main power supply causes multiple conveying lines to shut down at the same time, or an urgent insertion of a high-priority order causes global priority adjustment), more than 50% of the allocated pairs are affected, it is determined as "global abnormality", and global re-planning (re-executing S2-S5) is started.

[0178] Among them, as shown in Figure 4 The flow of local re-planning, exemplarily, specifically includes, (1) S401, affected subgraph construction: subgraph vertex screening: only keep the cone yarn nodes affected by the abnormality (such as the cone yarn with speed abnormality, the cone yarn that cannot be transmitted due to path failure) and "affected warping machine nodes" (such as the original allocated warping machine corresponding to the cone yarn, the standby warping machine that can accept the affected cone yarn), and exclude unaffected nodes (such as other normally transmitted cone yarns, normally hung warping machines).

[0179] Subgraph edge screening: based on the graph modeling logic of S2, only the affected nodes are reconstructed with directed edges, the screening condition is consistent with S4 (color number compatible, tension adapted, path reachable), but the "backup path" or "backup warping machine" is preferentially selected (such as Conveyor-05 failure, the path of Conveyor-06→Conveyor-08 is re-planned for the affected cone yarn).

[0180] (2) S402, subgraph edge scoring and matching: edge score updating: input the constructed subgraph into the pre-trained GNN model of S3 to recalculate the matching scores of all candidate edges in the subgraph, at this time the model will automatically increase the adaptation weight of "backup resources" (such as the score of the second standby warping machine (ID for WJ-09) will be adjusted according to the current idle state).

[0181] Subgraph matching solution: use the algorithm of S5 (Hungarian algorithm for small subgraph, minimum cost flow algorithm for large subgraph), solve the local optimal matching of "maximize total matching score" in the subgraph, and generate "local adjustment scheme" (such as reassigning the original yarn package TS-20240510-001 assigned to the warping machine one (ID WJ-08) to the warping machine two (ID WJ-09)).

[0182] (3) S403, adjustment instruction issuing and execution connection: only for the equipment involved in the local adjustment scheme to generate new control instructions (such as generating preparation instructions for the warping machine two (ID WJ-09) and generating transmission instructions for Conveyor-06), and the unaffected equipment continues to execute the original instructions, avoiding global instruction restart.

[0183] Set the adjustment connection time window (such as issuing new instructions within 2 minutes after the original instruction execution is interrupted), to ensure that the processing of the affected yarn package is not interrupted (such as switching to Conveyor-06 transmission within 2 minutes after the yarn package stops at Conveyor-05, without obvious waiting time).

[0184] Therefore, through the above local re-planning process, abnormal problems can be quickly solved without interfering with the normal scheduling, ensuring the continuity and stability of the warping workshop scheduling.

[0185] Figure 5 An intelligent scheduling management system 500 for a warping workshop is shown. The device embodiment corresponds to the method embodiment shown in Figure 1 The device embodiment corresponds to the method embodiment shown in

[0186] The bipartite graph construction module 502 is used to establish a bipartite graph with yarn task nodes and warping machine nodes as vertices, and to establish directed edges based on color number matching, path accessibility and tension constraints, and to generate feature vectors for nodes and edges.

[0187] The matching score module 503 is used to input the bipartite graph into a pre-trained graph neural network model, wherein the graph neural network model encodes and aggregates node and edge information through a node embedding layer and multiple message passing layers, and finally outputs a matching score representing the overall matching advantage for each directed edge through an edge score head.

[0188] The directed edge filtering module 504 is used to filter all the directed edges based on preset hard constraints, and to remove edges that do not meet color number compatibility, tension range or path accessibility.

[0189] The distribution scheme obtaining module 505 is configured to take the filtered directed edges and the matching scores thereof as inputs, construct and solve a weighted bipartite graph global matching problem aiming at maximizing the total matching score, so as to obtain an optimal distribution scheme of the cheese yarn tasks to the warpers.

[0190] The execution module 506 is configured to convert the distribution scheme into specific device control instructions and issue the same to a workshop execution system for execution.

[0191] The above merely provides a specific implementation of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which shall be encompassed in the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. An intelligent scheduling management method for a beaming workshop, characterized in that, Comprising the following steps: S1, obtaining a beaming workshop data, the beaming workshop data comprising a cone yarn task data, a beaming machine state data, a workshop topology and path information and a historical scheduling record; S2, establishing a bipartite graph with cone yarn task nodes and beaming machine nodes as vertices, establishing directed edges based on color number matching, path accessibility and tension constraints, and generating feature vectors for nodes and edges; S3, inputting the bipartite graph into a pre-trained graph neural network model, wherein the graph neural network model encodes and aggregates node and edge information through a node embedding layer and multiple message passing layers, and finally outputs a matching score representing the overall matching quality for each directed edge through an edge scoring head; S4, filtering all the directed edges based on pre-set hard constraints, and removing edges that do not meet color number compatibility, tension range or path accessibility; S5, taking the filtered directed edges and their matching scores as input, constructing and solving a weighted bipartite graph global matching problem with the goal of maximizing the total matching score, thereby obtaining the optimal cone yarn task to beaming machine allocation scheme; S6, converting the allocation scheme into specific device control instructions and issuing them to the workshop execution system for execution.

2. The intelligent scheduling management method for a beaming workshop according to claim 1, characterized in that, Further comprising preprocessing the beaming workshop data, including setting a sliding time window to collect time series data, normalizing continuous numerical features, numerically encoding or embedding discrete categorical features, and filling in missing data.

3. The method according to claim 1, characterized in that, The cone yarn task node contains normalized speed sequence summary, remaining cone length, color number embedding vector, tension level, priority and current path load features; the beaming machine node contains current machine speed, target color number embedding vector, allowed tension range, state code, priority and expected idle time features; the directed edge contains speed matching degree, path transmission delay estimation, historical pairing success rate, color number compatibility score and tension penalty value features.

4. The intelligent scheduling management method for a beaming workshop according to claim 1, characterized in that, Comprising: The bipartite graph is a time series dynamic graph that integrates the time series change characteristics of cone yarn tasks and beaming machines into the feature representation of nodes and edges by introducing time coding or using a sliding time window to capture the dynamic evolution of workshop state.

5. The method according to claim 1, characterized in that, The graph neural network model uses a bidirectional message passing mechanism, including forward message passing from cone yarn task nodes to beaming machine nodes, and reverse message passing from beaming machine nodes to cone yarn task nodes; message passing fuses the hidden state features of connected nodes and edge features, and uses attention mechanism to weight and aggregate messages from different neighbor nodes.

6. The method according to claim 1, characterized in that, The training of the graph neural network model uses a multi-objective joint loss function for supervised training; the joint loss function at least includes: a supervised learning loss based on historical successful pairing data, a ranking comparison loss for ensuring that high-quality pairing scores are higher than low-quality pairing, and a constraint-aware loss for imposing a penalty on pairing that violates tension or path capacity constraints.

7. The method according to claim 1, characterized in that, To solve the weighted bipartite graph global matching problem, the Hungarian algorithm or the minimum cost flow algorithm based on the maximum flow is used to obtain the global optimal matching under the constraints of satisfying the maximum allocation of each cone yarn task once and the task slot capacity of each warping machine.

8. The method according to claim 1, characterized in that, During execution, the system state is monitored in real time and compared with the prediction of the graph neural network. When the deviation exceeds the threshold, a local or global re-planning process is triggered to re-execute steps S2 to S5 and generate a new scheduling scheme.

9. The method according to claim 8, characterized in that, The re-planning process is local re-planning. When an anomaly is detected, only the subgraph composed of the cone yarn task nodes and the warping machine nodes affected by the anomaly event is re-scored and matched, while the scheduling scheme of the unaffected part of the system remains unchanged.

10. An intelligent scheduling management system for a beaming workshop, characterized in that, Comprise: A warping shop data acquisition module for acquiring warping shop data, including cone yarn task data, warping machine state data, shop topology and path information, and historical scheduling records; a bipartite graph construction module for establishing a bipartite graph with cone yarn task nodes and warping machine nodes as vertices, and establishing directed edges based on color matching, path accessibility and tension constraints, and generating feature vectors for nodes and edges; A matching score module for inputting the bipartite graph into a pre-trained graph neural network model, wherein the graph neural network model encodes and aggregates node and edge information through a node embedding layer and multiple message passing layers, and finally outputs a matching score representing the overall matching advantage and disadvantage for each directed edge through an edge scoring head; A directed edge filtering module for filtering all directed edges based on preset hard constraints, and removing edges that do not meet color compatibility, tension range or path accessibility; An allocation scheme acquisition module for inputting the filtered directed edges and their matching scores as input, constructing and solving a weighted bipartite graph global matching problem with the goal of maximizing the total matching score, and obtaining the optimal cone yarn task to warping machine allocation scheme; an execution module for converting the allocation scheme into specific device control instructions and issuing them to the shop execution system for execution.

Images